Abstracts

Rationale:
Resective surgery is an important treatment option for patients with drug-resistant focal epilepsy (DRE) and relies on accurate localization of the epileptogenic zone (EZ), or brain region generating seizures. The gold standard for defining the EZ is localization to the resected area in patients who become seizure free following epilepsy surgery. Therefore, resection masks are critical to EZ localization. Widely available brain segmentation software may not work well in patients with significant brain abnormalities or post-operative patients. Manual creation of resection masks is time consuming and user dependent. Here, we propose automated resection mask creation based on the Classifier, a method that uses a template-free machine learning approach to classify brain tissue through multi-scale local image features. This method could increase accuracy and reproducibility compared to manual or other available approaches.




Methods:
Study participants consist of 5 healthy volunteers (HVs), 10 adult NIH patients and 10 Children’s National Hospital pediatric patients who underwent epilepsy surgery with available pre- and post-operative 3T MR images using Siemens and GE scanners, respectively. For each voxel in an image, we computed 3D rotationally invariant Gaussian, gradient magnitude, Laplacian and Hessian magnitude filters at 3 spatial scales (FWHM 2, 4, 8 mm) yielding a 13-feature signature. We created labeled training masks in 5 HVs using FreeSurfer cortical gray matter (GM), cerebral white matter (WM), and sMRIPrep-derived CSF to train a multinomial logistic regression model. We then applied the model to pre- and post-operative images using a winner-take-all classification system. Pre- and post-op images were co-registered linearly and non-linearly, and combined GM and WM masks were subtracted to obtain a pre- versus post-op difference mask. The pipeline optimizes the mask by choosing the largest cluster. We compared Classifier generated masks to a manually refined mask, an SPM-based difference pipeline used by Children’s National, and an sMRIprep-based pipeline (https://pypi.org/project/smriprep/).




Results:
For the NIH adult patients, compared to a manually created mask, we found an average Dice Similarity Coefficient (DSC) of about 0.97 for the Classifier-generated masks and a DSC of 0.40 for the sMRIprep masks, with an average sensitivity of 0.98 vs. 0.81, respectively. False positive rates were similar at approximately 0.05. Compared to manually refined masks in the 10 pediatric patients, the masks appear visually similar, with average DSC of the Classifier-generated masks of 0.98 compared to 0.60 using an SPM-based mask.




Conclusions:
Our approach demonstrates increased similarity to manually created resection masks when compared to the pre-post-operative difference image created using sMRIprep. We found similar results using T1 images obtained at two centers, in children and adults, and in temporal and extratemporal resections. Overall, these results suggest that the automated Classifier approach generates accurate and reproducible resection masks, offering an efficient alternative to manually drawn masks.




Funding: Funding was provided through the NIH Intramural Research Program.